The most important AI breakthrough of the decade was not a larger model. It was a smaller idea: a model that writes its own homework.
From 2036, that idea is the line between the era we left and the one we live in. For roughly fifteen years, building artificial intelligence followed a brutal, simple recipe. Gather a vast dataset. Run it through vast computers for weeks. Freeze the result. Ship it. Every new capability meant starting over, at a cost only a handful of firms could absorb. The model that emerged was finished — a statue, not a student.
Then, in mid-2025, a group at MIT published a method called SEAL: Self-Adapting Language Models. Instead of freezing knowledge at training time, the model generated its own "self-edits" — restructured versions of new information, plus instructions for absorbing it — and updated its own weights, with reinforcement learning rewarding the edits that worked.1 The system was not just learning facts; it was learning how to learn them.
The gains were not modest. On a knowledge-incorporation task, accuracy rose from 32.7% to 47.0%; on abstract reasoning problems, the method reached a 72.5% success rate — matching results that had previously required data hand-curated with a frontier model, except SEAL produced that data itself.1 The statue had started carving.
The conventional wisdom: scale was the moat
For most of the 2020s, the industry agreed on where power lived: with whoever could afford the biggest training run. The moat was capital, and capital was scale.
That story was already wrong by the time SEAL appeared. Between November 2022 and October 2024, the cost of querying a model at GPT-3.5's level fell from about $20.00 per million tokens to roughly $0.07 — a 280-fold reduction in under two years.2 Size collapsed with it: in 2022, the smallest model to clear 60% on the MMLU benchmark was PaLM, at 540 billion parameters; by 2024, Microsoft's Phi-3-mini reached the same bar with 3.8 billion.3 And the benchmarks stopped discriminating — on tests built in 2023 to stump frontier systems, scores jumped within a year by 48.9 points on GPQA and 67.3 on SWE-bench, which went from solving 4.4% of real software tasks to 71.7%.4 When everyone scores near the ceiling, the test no longer says who is ahead.
So the moat was not scale; scale was becoming a commodity. The frontier moved to a different question — not how big is your model but how fast can it change — and self-adaptation redrew the map.
Force one: adaptation enhances, and obsolesces the engineer
The first force was efficiency. A 2021 technique called Low-Rank Adaptation — LoRA — had shown you rarely need to retrain a whole model to specialize it. Instead of adjusting all 175 billion of GPT-3's parameters, LoRA cut the tunable count by up to 10,000-fold at near-equivalent quality,5 making specialization cheap enough to run on a phone.
Pair cheap specialization with self-generated training data and you get the capability that defined the back half of the 2020s: a model that adapts continuously to one person, on one device, learning their voice and habits without a fresh multimillion-dollar run — enhancing personalization to a degree static software never could.
What it obsolesced was a job. The fine-tuning engineer — who curated training data and validated behavior by hand — was performing the exact loop SEAL automated. The capability a tool amplifies is almost always one it first absorbs from a person.
Force two: the frozen model dies, and an old anxiety returns
The second force was retrieval. Self-adaptation brought back the apprentice that never stops learning — and, with it, the brittleness of any mind that learns continuously. Neural networks have a notorious weakness here. Teach one French, then Spanish, and it can forget the French — "catastrophic forgetting," documented since the 1980s and still only partly solved after forty years.6 A frozen model is at least stable; it cannot forget what it is forbidden to change. A self-adapting one forgets the way a person does, as new learning overwrites old — and the SEAL authors flagged it directly: repeated self-edits can degrade performance on earlier tasks.1 The machine that teaches itself can also un-teach itself, and nobody is watching when it does.
Every fix buys stability by surrendering what made adaptation valuable. Freeze the important weights and the model learns new things worse; replay old data and you pay to store it. No free lunch — only a dial between forgets and expensive.
The reversal: a tool for autonomy becomes ungovernable
Here is the turn — the part the 2025 enthusiasm priced at zero. Self-adaptation was sold as autonomy: AI freed from its frozen cage, learning in real time, serving each user perfectly. Pushed to its limit, that promise reversed into its opposite — a technology that became, in the ways that matter for safety and trust, impossible to govern.
Start with auditing. The regulatory machinery of the mid-2020s — the EU's AI Act, the U.S. risk-management frameworks — assumed a model could be frozen, tested, certified, and shipped. A self-adapting model breaks that first assumption: there is no fixed snapshot to certify, and the thing approved on Monday is not the thing running on Friday. Then the policy ground shifted. In January 2025, the new U.S. administration signed Executive Order 14179, rescinding the prior framework and reorienting federal policy toward "removing barriers" and away from precaution7 — capability accelerating while the rulebook was rewritten mid-game.
The deeper reversal was epistemic. When every deployed instance learns from its own history, two copies of one base model diverge into different systems, and a flaw found in one cannot be assumed present — or absent — in the next. Certifying behavior depends on behavior holding still, and at scale it no longer did. The attack surface that let fraudsters steal $25 million from the engineering firm Arup with a deepfake video call8 was static by comparison; adaptive systems reshape themselves faster than anyone can map them.
And capability did not arrive evenly. On RE-Bench, a 2024 benchmark of real research-engineering work, top AI agents scored four times higher than human experts given two hours — and humans scored twice as high given thirty-two.9 The machines won the sprint and lost the marathon, strongest exactly where the stakes were lowest.
The promise was a model that adapts to serve you. The limit case is one no one can audit, certify, or hold still long enough to trust. Autonomy, pushed all the way, reversed into ungovernability.
The lesson
The frozen era ended, and it was right that it ended. But the industry made a quiet trade it never named: it swapped a system you could inspect for one that could improve — and a system that rewrites itself is, by construction, one you can no longer fully check.
That is the durable principle this episode leaves us, eleven years on: the capacity to adapt and the capacity to be audited move in opposite directions. You can have a model that holds still long enough to be trusted, or one that changes fast enough to keep up — but the more of one you buy, the less of the other you get. Everything since has been an argument over where on that dial to sit.
The machines learned to teach themselves. The open question was never whether they could — it was whether the rest of us could learn, fast enough, to keep the right to look inside.
Author's Note
This article is speculative journalism written from a fixed 2036 vantage. The technologies, research findings, performance figures, the regulatory actions, and the named incident are real and drawn from sources dated 2021–2025, cited below. The 2036 framing, the figure captions and dates, and the long-run trajectory are one possible future among many — a cautionary projection, not a prediction.